Research on Interpretability of Malware Detection Models for Image-Based Analysis

In recent years, image analysis-based malware detection methods built by convolutional neural networks have received much attention, and their models' intelligence has gradually been valued. However, the ‘black-box’ nature of the malware detection model based on convolutional neural networks makes it impossible for the model designer to explain why the model can make a specific decision. An interpretability study of our proposed malware detection method based on image analysis and generative adversarial networks is conducted to address this issue. First, we generate gray and Markov images of malware and feed them into the target model for feature extraction and identification, eliminating the need to extract malware features manually. Second, we generate heatmaps and cumulative heatmaps of the same malware family based on the gradient-weighted class activation mapping algorithm, and by observing the cumulative heatmaps, the key features of malware can be further interpreted and understood. In addition, we calculate the average heatmap similarity of malware families based on the cosine similarity algorithm, which helps to verify the reasonableness of the target model's prediction by comparing the average heatmap similarity results of the same families with different families. The experimental results show that in the BIG 2015 and CCF Malware datasets, the average heatmap similarity generated by the malware families is higher than the average heatmap similarity of the mixed families, and the similarity of the same families is higher than 85%, which demonstrates that the MadInG model can classify the malware correctly, and its prediction results are reasonable.